seedance2-skill: a prompt writing skill for Seedance 2.0 video prompts

A skill file rather than a prompt library

The repository is small enough to describe completely. Its tree has seven entries: `.gitignore`, `LICENSE`, `README.md`, `README-zh.md`, `SKILL.md`, an `assets/` directory holding the thumbnail image, and a `zh/` directory with the Chinese version of the skill. There is no `package.json`, no build script, no test suite and no release history. What you install is Markdown meant to be read by an AI agent, not a library you import.

The repository description calls it a skill to create best prompts for generating videos with seedance2.0, and the README expands on that: agent skills for writing effective video generation prompts for Jimeng Seedance 2.0, which it identifies as ByteDance's multimodal AI video generation model. It names Claude Code, Cursor and Cline as the agents it works with, and links to agentskills.io as the format it follows. The four repository topics, `prompt`, `seedance`, `seedance-2` and `video-prompt`, describe a text resource rather than a program.

That distinction changes what an evaluation looks like. There is no API wrapper, no prompt database and no cost to run, which also means there is no runtime behaviour to inspect. The repository is under the MIT license, it is not archived, and the last push was on 2026-09-19. The listed homepage is `seedance2.help`, a separate site from ByteDance's own product page at jimeng.jianying.com.

Copying SKILL.md into the Claude skills directory

Installation is the one place where the README has real commands, and it offers two routes. The first is the manual copy, which the README recommends. You create the skills directory, then copy the English skill file and the Chinese one into it:

mkdir -p ~/.claude/skills

# English
cp SKILL.md ~/.claude/skills/seedance-prompt-en.md

# Chinese
cp zh/SKILL.md ~/.claude/skills/seedance-prompt-zh.md

The rename on the way in is the interesting part. The file becomes `seedance-prompt-en.md` rather than keeping its original name, which is what lets both language versions live side by side in a directory that normally expects one file per skill. If you only write prompts in one language, copying one of the two lines is enough.

After copying, the README's only instruction for use is to ask your AI agent to help write a Seedance 2.0 video prompt. There is no configuration file, no environment variable and no validation step, because the skill is prompt text and the agent picks it up from the skills directory.

The skills CLI as a one line install

The second route is a single command:

npx skills add dexhunter/seedance2-skill

This is the shorter path if you already trust the skills CLI that agentskills.io points to, and it avoids deciding yourself which of the two language files to install. The README presents the manual copy first and labels it recommended, which is a reasonable default: copying two files tells you exactly what lands on your disk, while the CLI is faster when you are installing several skills at once.

Neither route changes what the skill contains. Both end with the same files in the same directory, so the choice between them is about how much you want to see happen on your machine, not about features.

What the two skill files claim to cover

The README's description of the guide is a list of five areas: input constraints, `@` reference syntax, camera language, prompt structure patterns, and ready to use templates for ads, dramas, MVs, educational content and more. The first two are the ones specific to Seedance 2.0. Input constraints describe what the model will and will not accept in a prompt, and `@` reference syntax is how a prompt points at supplied reference media rather than describing it in words.

Camera language and prompt structure patterns are not unique to this model, and the README does not claim they are. Their value here is that they arrive pre-bundled with the model specific material instead of living in a separate prompting guide. The template list maps to categories that show up in the case studies the README cites, which suggests the templates were built from real production work rather than written generically.

What the README does not do is reproduce any of it. Neither `SKILL.md` nor the README shows an example prompt, so a reader who wants to judge the guidance before installing has to install it. The tree also shows no `examples/` directory and no screenshots beyond the thumbnail in `assets/`.

Where the README stops and ByteDance's wiki begins

The README is explicit that it is built on official ByteDance documentation, and it links two Lark wiki pages. The first is the Seedance 2.0 User Manual, described as covering parameters, interaction methods, multimodal capabilities and example prompts. The second is Seedance 2.0 Real-world Cases, described as covering drama production, e-commerce ads, dance imitation, science education and AI MVs.

Those two links are the honest limit of this repository. Anything about the model's actual parameters, supported durations, resolution choices or moderation rules lives in the User Manual, not here, and this project deliberately stays upstream of that detail. The companion site at `seedance2.help` appears to be the author's own surface for the same subject.

The split is sensible for a portable skill, since parameters change with the hosted model and a skill file that copied them would go out of date. It does mean the skill cannot settle a question about a specific setting. When an agent following this skill gives you something the model rejects, the parameter list in the User Manual is where the answer lives.

Where this belongs against the alternatives

The straightforward comparison is against writing prompts yourself or reading ByteDance's User Manual directly. Reading the manual gives you the authoritative parameters and nothing else. This repository adds the layer above that: camera language, structural patterns and category templates, packaged so an agent applies them without being asked. Neither replaces the other, and the README's own source list makes that clear by crediting the manual rather than restating it.

The other alternative is a general prompt engineering guide. Those tend to be prose rather than something an agent loads as a skill, so they depend on the user pasting advice into a conversation each time. A skill file in `~/.claude/skills` is picked up without that step, which is the practical argument for this format.

The limits are equally plain. There is no versioning, so an update arrives as a new push to `main` with no tag to compare against, and nothing tells you whether the skill has been revised for a model update. There is also no mechanism for reporting a bad prompt back to the author beyond the repository's issue tracker. For a prompt resource under an MIT license that is a workable arrangement, but it is a different maintenance story from a released tool.

Editorial conclusion

seedance2-skill is worth a few minutes if you generate Seedance 2.0 clips regularly and your agent keeps producing prompts that miss the model's input constraints, since the whole repository is one English skill file and one Chinese skill file under an MIT license. It is the wrong tool if you need prompt templates for a different video model, if you want version history, or if you would rather read ByteDance's own parameter documentation directly, because this project points at those pages rather than reproducing them. There are no releases and no build, so there is nothing to upgrade and nothing to pin. Start by copying SKILL.md into your agent's skills directory and asking for one prompt on a real shot you need, then read the User Manual page the README cites when the agent gets a parameter wrong.

Frequently asked questions

Official sources

1. dexhunter/seedance2-skill on GitHub
2. Issues
3. License: MIT
4. Project website
5. README